Human-AI Interaction, Chatbots & Social Norms
Date: Monday 10 Nov 2025
Room: S/3.20 Turing Suite
| Time | Presentation | Abstract | Authors | Contribution type |
|---|---|---|---|---|
| 11:00-11:20 | From Queries to Prompts: Comparing User Experience in Generative AI Tools and Search Engines | Recent advances in large language models (LLMs) and the rise of Generative Artificial Intelligence (GenAI) tools, such as ChatGPT and Copilot, are ushering in a significant shift in the way people interact with information-seeking systems. This study presents a mixed-methods investigation aimed at comparing user experiences of GenAI tools and Conventional Search Engines (CSEs). Twenty-four participants completed fact-finding and browsing tasks using both types of tools. Quantitative data was gathered using Tobii Fusion eye tracking device and a paper-based NASA-TLX survey, while qualitative data was gathered through semi-structured interviews after task completion. Results revealed that GenAI prompts were significantly longer and more conversational, and GenAI tools imposed higher cognitive load during fact-finding, but less cognitive load during browsing tasks. Qualitative findings indicated that users value GenAI for abstract, creative and personalised tasks, but expressed concerns over accuracy, trust, and data privacy. This study expands the limited body of research on comparing user behaviour and experiences when seeking information using CSEs and GenAI tools. It offers a novel contribution by identifying differences in cognitive load associated with completing different task types across the different tool types, highlighting patterns in GenAI interaction behaviours, while also identifying the factors that influence user preferences, perceptions, and overall experience of GenAI tools. The paper concludes with a discussion of the implications of these findings and provides recommendations for designing GenAI tools to enhance user experience. | Misbahu Zubair, Muhammad Alhassan and Farid Bello | Research |
| 11:20-11:40 | “Chattable” Avatars: Using LLMs to Power Visitor Engagement with Historical Persons | Cultural Heritage institutions such as Galleries, Libraries, Archives, and Museums (GLAMs) are tasked with preserving our history and heritage for future generations while engaging new audiences with an appetite for said engagement to be increasingly digital and interactive. To explore how advances in natural language processing, particularly large language models (LLMs), may help GLAMs in their mission, we designed a prototype ‘Chattable’ avatar, a 3D high-polygon animated character which visitors can talk to and interact with. We report the design of our avatar, and a workshop we conducted with curators and staff from a GLAM institution, to understand the problems, requirements, and opportunities LLMs present in the cultural heritage space. We present results from a qualitative analysis of our workshop highlighting themes such as trust, authority, social experience, and location, finding LLMs may be more suited to deployments focused on non-factual data dissemination. We conclude with implications for GLAMs and suggestions for future research to realise how best to integrate GenerativeAI like LLMs into the GLAM space. | Zhuoling Jiang, Yipeng Qin and Daniel J. Finnegan | Research |
| 11:40-11:50 | Towards Explainable User Interfaces | This paper argues that user interfaces need to be explainable whether or not they contain artificial intelligence components. Even with the best design, complex applications often leave users confused; this is exacerbated on small touchscreens, where small slips can lead to markedly different outcomes and when notifications or intelligent agents may autonomously change the interface. This can be disorienting even for the most tech savvy user, but doubly so for those less confident or with motor-control issues. We are often left asking “what just happened?” or “how can I do this again?”. We need explainable user interfaces. | Alan Dix, Tommaso Turchi and Ben Wilson | Futures |
| 11:50-12:00 | Exploring User Acceptance of a Fintech Chatbot powered by LLMs among Older Adults | Many banks have shut their physical offices and shifted to online banking. Many will benefit from this transition, but older adults may be left behind due to their traditional views on money and technology. Without proper consideration of their needs and preferences, older adults may become marginalized in this digitalization process. In this paper, we investigate why older adults may or may not be comfortable with the technology-enabled ways of banking and their acceptance of a fintech chatbot. Towards this, we built a chatbot specifically designed to help older adults complete financial transactions such as money transfers, checking balances and managing their pensions. We developed this chatbot using human-centred design principles and inclusive and accessibility design methods. We incorporated Large Language Models (LLMs) into our prototype to enable users to engage in casual conversations with the chatbot to mimic chats with human bank advisors in physical venues. To assess the effectiveness of the prototype, we carried out user studies involving older adults. We analysed the study’s results to identify the factors that contribute to the trust and acceptance of the chatbot and in online banking in general. We further extrapolate these findings to provide recommendations for designing fintech chatbots for older adults. | Swaroop Panda, Farkhandah Komal, Effie Lai-Chong Law, Syed Murad, Zhongtian Sun, Ben Summerill, Delali Konu and Thuy-Vy Nguyen | Late Breaking Work |
| 12:00-12:10 | Social Norms, Social AI : Investigating the Effects of AI (Im)politeness and Gender on User Perception | The inherently social functionality of conversational AI systems requires attention – how users interact with these systems has the potential to both reflect and implicate imbalanced social norms. Such interactions are the result of specific design choices, including the ‘gendering’ of many conversational AI, which is typically based in stereotyped characterisations and thus poses specific risks for external gender relations. To analyse the extent to which these design choices affect user perception, we conduct a 2x2 between-subjects experiment manipulating the language style and ‘gender’ of four conversational AI systems and initiating unstructured human-AI interaction with participants. The subsequent participant ratings of the conversational AI provide insight into how design choices and users’ expectations intersect. The results found a significant overall preference for polite AI but no overall effect of AI gender on participant perception, although isolating participant gender highlighted gender-specific characterisations and preferences for conversational AI. | Aoife O’Driscoll and Alan Blackwell | Late Breaking Work |